Import & Export H2O Frames from/to S3

Sparkling Water car read and write H2O frames from and to S3. Several configuration steps are required.

Specify the AWS Dependencies

In order to enable support for S3A/S3N, we need to start Sparkling Water with the following extra dependencies:

  • org.apache.hadoop:hadoop-aws:2.7.3
  • spark.jars.packages com.amazonaws:aws-java-sdk:1.7.4

For production environments, we advice to download these jars and add them on your Spark path manually by copying them to $SPARK_HOME/jars directory.

Additionally, we can also use the --packages option when starting Sparkling Water as:

./bin/sparkling-shell --packages spark.jars.packages com.amazonaws:aws-java-sdk:1.7.4,org.apache.hadoop:hadoop-aws:2.7.3

The same holds for PySparkling. We can also add the following line to the spark-defaults.conf file:

spark.jars.packages com.amazonaws:aws-java-sdk:1.7.4,org.apache.hadoop:hadoop-aws:2.7.3

This line ensures that we don’t need to specify the --packages option all the time.

Configuring S3A

In order to read and write to S3A, please add the following lines to the spark-defaults.conf file in the $SPARK_HOME/conf directory:

spark.hadoop.fs.s3a.impl                org.apache.hadoop.fs.s3a.S3AFileSystem
spark.hadoop.fs.s3a.access.key          {{AWS_ACCESS_KEY}}
spark.hadoop.fs.s3a.secret.key          {{AWS_SECRET_KEY}}

where {{AWS_ACCESS_KEY}} should be substituted by AWS access key and {{AWS_SECRET_KEY}} by AWS secret key.

Configuring S3N

In order to read and write to S3N, please add the following lines to the spark-defaults.conf file in the $SPARK_HOME/conf directory:

spark.hadoop.fs.s3n.impl                org.apache.hadoop.fs.s3native.NativeS3FileSystem
spark.hadoop.fs.s3n.awsAccessKeyId      {{AWS_ACCESS_KEY}}
spark.hadoop.fs.s3n.awsSecretAccessKey  {{AWS_SECRET_KEY}}

where {{AWS_ACCESS_KEY}} should be substituted by AWS access key and {{AWS_SECRET_KEY}} by AWS secret key.

Sparkling Water Example Code

When you configured Sparkling Water as explained above, please start Sparkling Shell as

./bin/sparkling-shell

Next, let’s start H2OContext. This context brings H2O support into Spark environment:

import org.apache.spark.h2o._
val hc = H2OContext.getOrCreate(spark)

Finally, read the data:

import java.net.URI
val fr = new H2OFrame(new URI("s3n://data.h2o.ai/h2o-open-tour/2016-nyc/weather.csv"))

PySparkling Example Code

When you configured PySparkling as explained above, please start PySparkling as

./bin/pysparkling

Next, let’s start H2OContext. This context brings H2O support into Spark environment:

from pysparkling import *
hc = H2OContext.getOrCreate(spark)

Finally, read the data:

fr = h2o.import_file("s3n://data.h2o.ai/h2o-open-tour/2016-nyc/weather.csv")

In PySparkling, you can also export the file to S3 as:

h2o.export_file("s3n://path/to/target/location")